Comparative Evaluation of Physics-informed, Hamiltonian, and Lagrangian Neural Networks for 2D Single Target Tracking
Chang-Ho Kang, Sunyoung Kim · Journal of Institute of Control Robotics and Systems · 2024
This study uses 2D robot simulation datasets to conduct a comprehensive comparative analysis between traditional nonlinear filtering methods and physics-based neural network models, specifically physics-informed neural networks, Hamiltonian neural networks, and Lagrangian neural networks. The primary focus is on assessing the performance of these models in accurately estimating the position and velocity of a simulated robot moving along predefined and random trajectories. The diversified dataset includes circular, elliptical, and sinusoidal pathways to evaluate the models’ learning and generalization abilities. The findings highlight the superior capabilities of physics-based neural networks in capturing the dynamics of physical systems, particularly under complex physical conditions and uncertainties. The study shows the potential benefits of these models over traditional nonlinear filtering methods, taking a big step toward their use in real-world applications.